Chinese Named Entity Recognition of Non-Performing Assets Based on BBC: BERT-BiLSTM-CRF

In the post-epidemic era, the number of non-performing assets and entities has been growing rapidly, leading to a surge in the number of related entities. However, the traditional method of constructing a dictionary of non-performing assets of enterprises using the identification method of BERT-CRF does not contribute significantly to the accuracy. In response to this problem, this paper is a proposal for a new type of model named BERT-BiLSTM-CRF(BBC) which provides Named Entity Recognition (NER) for non-performing assets by combining the pre-trained BERT language model and BiLSTM-CRF. Compared to the static word vector model, BERT is able to generate dynamic word vectors based on the context, which improves the accuracy of the semantic coding together with the accuracy of the NER task for non-performing assets. The word vectors obtained from the BERT model are fed into BiLSTM model for further training, and then CRF corrects the model's output to produce the labeled sequence with the highest score. The proposed model achieved an F1 score of 0.936, which represents a 6.727% improvement and a 6.120% accuracy increase in comparison with the baseline model.